EP1142229B1

EP1142229B1 - Distributed hierarchical scheduling and arbitration for bandwidth allocation

Info

Publication number: EP1142229B1
Application number: EP99973516A
Authority: EP
Inventors: Marek Stephen Piekarski; Ian David Johnson
Original assignee: Xyratex Technology Ltd
Current assignee: Seagate Systems UK Ltd
Priority date: 1998-12-22
Filing date: 1999-12-01
Publication date: 2006-07-26
Anticipated expiration: 2019-12-01
Also published as: EP1142229A1; DE69932533D1; CA2353622A1; GB9828143D0; ATE334534T1; AU1398900A; JP2002533995A; CN1338168A; WO2000038376A1; US7099355B1; KR20010086086A

Abstract

According to the present invention there is provided a scheduling and arbitration process for use in a digital data switching arrangement of the type in which a central switch under the direction of a master control provides the cross-connections between a number of high-bandwidth ports to which are connected on the ingress side of the central switch a number of ingress multiplexers, one for each high-bandwidth input port and on the egress side a number of egress multiplexers, one for each high-bandwidth output port, each ingress multiplexer including a set of N input queues serving N low-bandwidth data sources and a set of M virtual output queues serving M low-bandwidth output data sources, characterized in that the scheduling and arbitration arrangement includes three bandwidth allocation tables, an ingress port table associated with the input queues and having NxM entries each arranged to define the bandwidth for a particular virtual output queue, an egress port table associated with the virtual output queues and having M entries each arranged to define the bandwidth allocation of a high-bandwidth port of the central switch to a virtual output queue and a central allocation table located in the master control and having (MxN)<SUP>2 </SUP>entries each of which specifies the weights allocated to each possible connection through the central switch.

Description

The present invention relates to data switching systems and is more particularly concerned with the scheduling and arbitration arrangements for such systems.
The continual growth of demand for manageable bandwidth in networks requires the development of new techniques in switch design which decouples the complexity of control from the scale of the port count and aggregate bandwidth. This invention describes a switch architecture and a set of methods which provide the means by which switches of arbitrary size may be constructed whilst maintaining the ability to allocate guaranteed bandwidth to each possible connection through the switch. A digital switch is used to route data streams from a set of source components to a set of destination components. A cell-based switch operates on data which is packetised into streams of equal size cells. In a large switch the routing functions may be implemented hierarchically, that is sets of lower bandwidth ports are aggregated into a smaller number of higher bandwidth ports which are then interconnected in a central switch.
The article entitled "Traffic Control in ATM Switches with Large Buffers" published in ICT Specialists Seminar, NL, Leidschendam, KPM Research, Volume Seminar 9, 1995, pages 45 to 60 by Wallmeier et al discloses a method and apparatus for supporting statistical multiplexing in an ATM network. The switch architecture disclosed relies on the use of statistical multiplexing units.
It is an object of the present invention to provide a bandwidth allocation arrangement which may be used in such a hierarchical switch.
According to the present invention there is provided a method of scheduling the passage of data cells from M low-bandwidth data sources to M low-bandwidth data destinations, said method being performed by a data switching apparatus including: M/N ingress multiplexers, each arranged to receive data cells from a respective set of N said low-bandwidth data sources, M/N egress multiplexers, each arranged to transmit data cells to a respective set ofN said low-bandwidth data destinations, a master control unit, and a central switch having M/N high bandwidth input ports arranged to receive data cells from respective said ingress multiplexers, and M/N high-bandwidth output ports arranged to transmit data cells to respective said egress multiplexers, the central switch selectively interconnecting the input ports and output ports, under the direction of the master control unit, the method including: each said ingress multiplex maintaining N input queues for queuing data cells received from the N respective said data sources, and maintaining M virtual output queues for queuing data cells directed to respective said data destinations; and the method being characterised in that: each ingress multiplexer further maintains a respective ingress port table, each ingress port table having NxM entries, each entry corresponding to a respective combination of a said data source for that ingress port and a said data destination, each ingress multiplexer transfers data cells from said input queues to said virtual output queues with a relative frequency according to value of the corresponding entry of the ingress port table; each ingress multiplexer further maintains a respective egress port table, the egress port table having M entries, each entry corresponding to a respective said data destination, each ingress multiplexer transfers data cells from said virtual output queues to said respective input ports of the central switch with a relative frequency according to the value of the corresponding entry of the egress port table; the master control unit maintains a central allocation table having (M/N)² entries, each corresponding to a respective combination of an input ort and an output port, and the master control unit controls the central switch to interconnect pairs of said input ports and output ports with a relative frequency according to the value of the corresponding entry of the central allocation table; whereby said ingress port tables, egress port tables and central allocation table together determine the bandwidth through the digital data switching apparatus from each said data source to each said data destination.
According to a second aspect of the present invention a digital data switching apparatus for transmitting data from M low-bandwidth data sources to M low-bandwidth data destinations, the apparatus including: M/N ingress multiplexers for receiving data cells from respective sets ofN said low-bandwidth data sources, M/N egress multiplexers for transmitting data cells to respective sets of N said low-bandwidth data destinations, a master control unit, and a central switch having M/N high-bandwidth input ports arranged to receive data cells from respective said ingress multiplexers, and M/N high-bandwidth output ports arranged to transmit data cells to respective said egress multiplexers, the central switch being arranged selectively to interconnect the input ports and output ports, under the direction of the master control unit, each said ingress multiplexer being arranged to maintain N input queues for queuing data cells received from respective said data sources, and to maintain M virtual output queues for queuing data cells directed to respective said data destinations; characterised in that: each ingress multiplexer is arranged to maintain a respective ingress port table, each ingress port table having NxM entries, each entry corresponding to a respective combination of a said data source and a said data destination, and each ingress multiplexer is arranged to transfer data cells from said input queues to said virtual output queues with a relative frequency according to value of the corresponding entry of the ingress port table; each ingress multiplexer is arranged to maintain a respective egress port table, the egress port table having M entries, each corresponding to a respective said data destination, and each ingress multiplier is arranged to transfer data cells from said virtual output queues to said respective input ports of the central switch with a relative frequency according to value of the corresponding entry of the egress port table, and the master control unit is arranged to maintain a central allocation table having (M/N)² entries, each corresponding to a respective combination of an input port and an output port, and the master control unit controls the central switch to interconnect pairs of said input ports and output ports with a relative frequency according to the value of the corresponding entry of the central allocation table; whereby said ingress port tables, egress port tables and central allocation table together determine the bandwidth through the digital data switching apparatus from each said data source to each said data destination.
According to a feature of the invention there is provided a scheduling and arbitration process in which the scheduling of the input queues is performed in accordance with an N-way weighted round robin.
According to a further feature of the invention there is provided an implementation of the N-way weighted round robin by an N.(2^w-1)-way unweighted round robin where w is the number of bits defining a weight using a list constructed by interleaving N words of (2^w-1) bits each, with w _n 1's in a word, where w _n is the weight of the queue n.
The invention, together with its various features, will be more readily understood from the following description of one embodiment, which should be read in conjunction with the accompanying drawings. In the drawings:-

Figure 1 shows a simplified form of a data switch,
Figure 2 shows an egress multiplexer,
Figure 3 shows the weighted round robin arbiter for use in the egress multiplexer,
Figure 4 shows the partitioning of the round robin arbiter,
Figure 5 shows the operation of the round robin arbiter,
Figure 6 shows the allocations for a 4-port interconnect with 3 bit weights, and
Figure 7 shows a block diagram of a small switch based on the principles of the invention.

Referring now to Figure 1, this shows a schematic diagram of a hierarchical switch. The central interconnect 1 provides the cross-connections between a number of high-bandwidth ports. A set of multiplexers 2 on the ingress side and demultiplexers 3 on the egress side provides the aggregation function between the low and high-bandwidth ports. The low bandwidth ports provide connections from the switch to the data sources 4 on the ingress side and the data destinations 5 on the egress side. In practice, a switch is required to support full-duplex ports, so that an ingress multiplexer and its corresponding demultiplexer may be considered a single full-duplex device which will be hereafter termed a "router" Typically the data switch may be of the type disclosed in our patent application WO 038375 published 29/06/2000.
It should be noted that the central interconnect 1 may itself be a hierarchical switch, that is the methods described may be applied to switches with an arbitrary number of hierarchical levels.
The aim of these methods is to provide a mechanism whereby the data stream from the switch to a particular destination, which comprises a sequence of cells interleaved from various data sources, may be controlled such that predetermined proportions of its bandwidth are guaranteed to cells from each data source.
Figure 2 shows the architecture of an ingress multiplexer. An ingress multiplexer receives a set of data streams from the data sources via a set of low-bandwidth input ports. Each data stream is a sequence of equal size cells (that is, an equal number of bits of data). A set of N low-bandwidth ports 21 each fills one of the N input queues 22. An ingress control unit 23 extracts the destination address from each of the cells in the input queues and transfers them into a set of M virtual output queues 24. There is one virtual output queue for each low-bandwidth output port in the switch. The ingress multiplexer also contains an interconnect link control unit 25 which implements this function by scheduling cells from the virtual output queues 24 across the high-bandwidth link 26 to the central interconnect 1 according to an M-entry egress table 28.
In addition to the data flow indicated by the arrows in Figure 1, there is also a flow of backpressure or flow-control information associated with each of the data flows. This control flow is indicated in Figure 2 by dashed arrows. The ingress multiplexer contains an NxM-entry ingress port table 27, which defines how its bandwidth to a particular egress port (via a particular virtual output queue) is distributed across the input ports. This table is used by the ingress control unit 23 to determine when (and to what degree) to exert backpressure to the data source resolved down to an individual virtual output queue.
The ingress multiplexer 2 of Figure 1 sends control information to the central interconnect 1 indicating the state of the virtual output queues in the form of "connection requests". The central interconnect responds with a sequence of connections which it will establish between the ingress and egress routers. These are "connection grants". The ingress multiplexer 2 must now allocate the bandwidth to each egress demultiplexer 3 provided by the central interconnect 1 across the virtual output queues associated with each egress demultiplexer.
The deterministic scheduling function of the interconnect link control unit 25 may be defined as a weighted round robin (WRR) arbiter. The interconnect link control unit 25 receives a connection grant to a particular egress demultiplexer 3 from the central interconnect 1 and must select one of the N virtual output queues associated with that egress demultiplexer . This may be implemented by expanding the N-way WRR shown in Figure 3a) into an (N.(2^W-1)) - way unweighted round robin as shown in Figure 3b), where W equals the number of bits necessary to define the weight, such that if a queue has a weight of w, then it is represented as (w-1) entries in the unweighted round robin list. For example, with 4-bit weights, a 4-way weighted round robin expands to a 60-way unweighted round robin.
In order to optimise the service intervals to the queues under all weighting conditions, the entries in the unweighted round robin list are distributed such that for each weight the entries are an equal number of steps apart plus or minus one step. Table 1 below shows an example of such an arrangement of 3-bit weights:

w _n e _n

1 1000000

2 1000100

3 1001010

4 1010101

5 1011011

6 1110111

7 1111111
In the system described, the arbiter must select one of the nine queues with 4-bit weights, that is 8 virtual output queues as described above and a multicast queue. This expands to a 135-entry unweighted round robin. The implementation of a large unweighted round robin arbiter may be achieved without resorting to a slow iterative shift-and-test method by the technique of "divide and conquer", that is the 135-entry round robin is segmented into 9 sections of 16-entry round robins, each of which may be implemented efficiently with combinational logic (9 x 16 provides up to 144 entries, so that the multicast queue of up to 24 entries may actually be allocated more bandwidth than an individual unicast queue of up to 15 entries).
Figure 4 illustrates the partitioning of the round robin arbiter. The sorter 41 separates the request vector V (144 bits) into 9 sections of 16-bit vectors , v0 to v8. It also creates nine pointers p0 to p8 for each of the 16-bit round robin blocks 42. The block which corresponds to the existing pointer (which has been saved in register 44) is given a "1" at the corresponding bit location, whilst the other blocks are given dummy pointers initialised to location zero. Each 16-bit round robin block now finds the next "1" in its input vector and outputs its location (g)whether it has to wrap round (w) and whether it has found a "1" in its its vector (f). A selector 43 is now able to identify the block which has found the "1" corresponding to the next "1" in the original 135-bit vector given a signal (s) from the sorter 41. This specifies which round robin block had the original pointer position. The selector 43 is itself a round robin function which may be implemented as a combinational logic function
"find the next block starting at s which has w=false and f=true (if not found, select s)".
Figure 5 shows an example of the above process, but for a smaller configuration for clarity. In the example, V = 12 bits, p = 4 bits, v0 - 2 = 2 bits and g0 - 2 = 2 bits. Figure 5 depicts the process performed by Figure 4 and at 51 defines the expanded current pointer (P) and the request vector (V) at 52. The sorter 41 produces segmented vectors (v) and segmented pointers (p) where the blocks marked * are dummies. The segmented results (g) of the round robin are shown at 55 whereas the results of the selector process 43 is shown at 56, defining the expanded next pointer (P).
The central interconnect 1 provides the cross-connect function in the switch. The bandwidth allocation in the central interconnect is defined by an (M/N)² -entry central allocation table, which specifies the weights allocated to each possible connection through the central interconnect (the central interconnect has M/N high-bandwidth ports). The central allocation table contains P² entries, where P=(M/N). Each entry w _ie defines the weights allocated to the connection from high-bandwidth port i to high-bandwidth port e. However, not all combinations of entries constitute a self-consistent set, that is the allocations as seen from the outputs could contradict the allocations as seen from the inputs. A set of allocations is only self-consistent if the sums of weights at each output and input are equal. Figure 6 shows a self-consistent set (a) and a non-self-consistent set (b) of allocation for a 4-port interconnect with 3-bit weights. Inputs are shown at IP and outputs at OP, with the sum designated as Σ. Assuming that the central allocation table has a self-consistent set of entries, it is possible to define the bandwidth allocation to a link between input port i and output port e with weight w _ie as p _ie, where: $p_{i e} = \frac{w_{i e}}{\sum_{n = 0}^{(p - 1)} w_{i n}}$
The egress port table defines how the bandwidth of a high-bandwidth port to the central interconnect 1 is allocated across the virtual output queues. There is no issue with self-consistence as all possible entries are self-consistent, so that the bandwidth allocation for a virtual output queue v with weight w _v is given by: $p_{f} = \frac{w_{v}}{\sum_{n = 0}^{(N - 1)} w_{n}}$

Similarly, the ingress port table entries give the bandwidth allocation of a virtual output queue to the ingress ports with weight w _f is given by: $p_{f} = \frac{w_{f}}{\sum_{n = 0}^{(N - 1)} w_{n}}$

Therefore the proportion of bandwidth at an egress port v allocated to an ingress port f is given by: $p_{f v} = p_{f} . p_{v} . p_{i e}$
In a switch which is required to maintain strict bandwidth allocation between ports (such as an ATM switch), the tables are set up via a switch management interface from a connection admission and control processor. When the connection admission and control processor has checked that it has the resources available in the switch to satisfy the connection request, then it can modify the ingress port table, the egress port table and the central allocation table to reflect the new distribution of traffic through the switch.
In contrast, a switch may be required to provide a "best effort" service. In this case the table entries are derived from a number of local parameters. Two such parameters are the length l _v of the virtual output queue v and the urgency u _v of the virtual output queue. urgency is a parameter which is derived from the headers of the cells entering the queue from the ingress ports.
A switch may be implemented which can satisfy a range of requirements (including the two above) by defining a weighting function which "mixes" a number of scheduling parameters to generate the table entries in real time according to a set of "sensitivities" to length, urgency and pseudo-static bandwidth allocation. (s _l ,sw, ss). The requirement on the function are that it should be fast and efficient, since multiple instances occur in the critical path of a switch. In the system described the weighting function has the form: $w_{v} = {\frac{l_{v}^{}}{2^{(1 / s l)}} + \frac{p_{v}}{2^{(1 / s s)}} + \frac{u_{v}}{2^{(1 / s u)}}} \cdot (1 - b_{v})$
where

b _v is the backpressure applied from the egress multiplexer,
w _v is the weight of the queue as applied to the scheduler, and
p _v is a pseudo-static bandwidth allocation, such as an egress port table.

Despite the apparent complexity of this function, it may be implemented exclusively with an adder, multiplexers and small lookup tables , thus meeting the requirement for sped and efficiency. Features of this weighting function are that, for s _l= 1.0, s _s = 0.0 and s _u = 0.0, bandwidth is allocated locally purely on the basis of queue length, with a non-linear function , so that the switch always attempts to avoid queues overflowing. When s _l = 0.0, s _s = 1.0 and s _u = 0.0, bandwidth is allocated purely on the basis of pseudo-static allocations a described above. Finally, when s _l = 0.0, s _s = 1.0 and s _u = 0.5, bandwidth is allocated on the basis of pseudo-static allocation but a data source is allowed to "push" some data harder, when the demand arises, by setting the urgency bit in the appropriate cell headers.

Figure 7 is a block diagram of a small switch based on the above principles, showing the correct number of queues, tables and table entries. In Figure 7 there are two

ingress routers

71 and 72, a central cross-bar switch 73 and two egress routers 74 and 75. Each ingress router has two low-bandwidth input ports, A and B for router 71 and ports C and D for router 72. As mentioned previously, each ingress router has an ingress port table such as 77 for router 72 and an egress port table such as 78, whereby the central switch 73 has a central allocation table 79. Assuming that each low-bandwidth port may transport 1 Gbps of traffic, each high-bandwidth link may carry 2 Gbps and the switch is required to guarantee the follwing bandwidth allocations:

Flow bandwidth	Destination Port
(Gbps)	A	B	C	D
A	0.5	0.1	0.1	0.2
B	0.2	0.2	0.2	0.2
C	-	0.5	-	0.2
D	0.1	0.1	0.6	0.2

then the ingress port table such as 77, egress port table such as 78 and central allocation table 79 would be set up by the connection admission and control processor with the following 4- bit values (note here that there will be rounding errors due to the limited resolution of the 4-bit weights):

Ingress Port Table			Ingress Port Table
(in router 71)			(in router 72)
Source			Source
	A	B		C	D
A	15	6	A	0	3
B	3	6	B	15	3
C	3	6	C	0	15
D	6	6	D	6	5

Egress Port Table		Egress Port Table
(in router 7 1)		(in router 72)
Source		Source
	AB		CD
A	14	A	2
B	6	B	12
C	6	C	12
D	8	D	8

Central Allocation Table
	Destination Router
		AB	CD
Source	AB	15	10
	CD	10	15

Claims

A method of scheduling the passage of data cells from M low-bandwidth data sources (4) to M low-bandwidth data destinations (5), said method being performed by a data switching apparatus including:
M/N ingress multiplexers (2; 71, 72), each arranged to receive data cells from a respective set of N said low-bandwidth data sources (4),

M/N egress multiplexers (3; 74, 75), each arranged to transmit data cells to a respective set of N said low-bandwidth data destinations (5),

a master control unit, and

a central switch having M/N high-bandwidth input ports arranged to receive data cells from respective said ingress multiplexers (2; 71, 72), and M/N high-bandwidth output ports arranged to transmit data cells to respective said egress multiplexers (3; 74, 75), the central switch selectively interconnecting the input ports and output ports, under the direction of the master control unit,

the method including:
each said ingress multiplexer (2; 71, 72) maintaining N input queues (22) for queuing data cells received from the N respective said data sources, and maintaining M virtual output queues (24) for queuing data cells directed to respective said data destinations;

and the method being characterised in that:
each ingress multiplexer further maintains a respective ingress port table (27; 77), each ingress port table having NxM entries, each entry corresponding to a respective combination of a said data source for that ingress port and a said data destination,

each ingress multiplexer transfers data cells from said input queues to said virtual output queues with a relative frequency according to value of the corresponding entry of the ingress port table (27; 77);

each ingress multiplexer further maintains a respective egress port table (28; 78), the egress port table having M entries, each entry corresponding to a respective said data destination,

each ingress multiplexer transfers data cells from said virtual output queues to said respective input ports of the central switch with a relative frequency according to the value of the corresponding entry of the egress port table (28; 78);

the master control unit maintains a central allocation table (79) having (M/N)² entries, each corresponding to a respective combination of an input port and an output port, and

the master control unit controls the central switch to interconnect pairs of said input ports and output ports with a relative frequency according to the value of the corresponding entry of the central allocation table (79);

whereby said ingress port tables (27; 77), egress port tables (28; 78) and central allocation table (79) together determine the bandwidth through the digital data switching apparatus from each said data source (4) to each said data destination (5).
A method according to claim 1 in which each said ingress multiplexer (2; 71, 72), for each virtual output queue (24), transfers data cells to that virtual output queue from said input queues in accordance with a N-way weighted round robin, using N weights determined respectively by the N entries of the ingress port table for that virtual output queue.
A method according to claim 2 in which each weight is defined by a number of bits w, and the N-way weighted round robin for each virtual output queue is implemented by an N(2^w-1)-way unweighted round robin using a request vector list constructed by interleaving N words of (2^w-1) bits each, each word corresponding to a respective input queue and having a number of "1"s determined by the entry of the ingress port table for that input queue and that virtual output queue.
A method according to claim 3 in which the request vector list is separated into a plurality of round robin blocks, each corresponding to a respective input queue, a first round robin process being performed independently within each block, and a second round robin process being performed to make a selection among the blocks.
A method according to any preceding claim in which the ingress port table (27; 77), the egress port table (28; 78) and the central allocation table (79) are all programmed from an external source.
A method according to claim 5 in which the external source uses parameters characterizing the length of each virtual output queue and the urgency of each virtual output queue.
A method according to claim 6 in which the external source uses a set of sensitivities relating to length, urgency and pseudo-static bandwidth allocation.
A digital data switching apparatus for transmitting data from M low-bandwidth data sources (4) to M low-bandwidth data destinations (5), the apparatus including:
M/N ingress multiplexers (2; 71, 72) for receiving data cells from respective sets of N said low-bandwidth data sources (4),

M/N egress multiplexers (3; 74, 75) for transmitting data cells to respective sets of N said low-bandwidth data destinations (5), a master control unit, and

a central switch having M/N high-bandwidth input ports arranged to receive data cells from respective said ingress multiplexers (2; 71, 72), and M/N high-bandwidth output ports arranged to transmit data cells to respective said egress multiplexers (3; 74, 75), the central switch being arranged selectively to interconnect the input ports and output ports, under the direction of the master control unit,

each said ingress multiplexer (2; 71, 72) being arranged to maintain N input queues (22) for queuing data cells received from respective said data sources, and to maintain M virtual output queues (24) for queuing data cells directed to respective said data destinations;
characterised in that:
each ingress multiplexer (2; 71, 72) is arranged to maintain a respective ingress port table (27; 77), each ingress port table having NxM entries, each entry corresponding to a respective combination of a said data source and a said data destination, and each ingress multiplexer is arranged to transfer data cells from said input queues to said virtual output queues with a relative frequency according to value of the corresponding entry of the ingress port table;

each ingress multiplexer is arranged to maintain a respective egress port table (28; 78), the egress port table having M entries, each corresponding to a respective said data destination (5), and each ingress multiplier is arranged to transfer data cells from said virtual output queues to said respective input ports of the central switch with a relative frequency according to value of the corresponding entry of the egress port table (28; 78),

and the master control unit is arranged to maintain a central allocation table (79) having (M/N)² entries, each corresponding to a respective combination of an input port and an output port, and the master control unit controls the central switch to interconnect pairs of said input ports and output ports with a relative frequency according to the value of the corresponding entry of the central allocation table (79);

whereby said ingress port tables (27; 77), egress port tables (28; 78) and central allocation table (79) together determine the bandwidth through the digital data switching apparatus from each said data source (4) to each said data destination (5).
An apparatus according to claim 8 in which each said ingress multiplexer is arranged, for each virtual output queue (24), to transfer data cells to that virtual output queue from said input queues (22) in accordance with a N-way weighted round robin, using N weights determined respectively by the N entries of the ingress port table (27; 77) for that virtual output queue.
A apparatus according to claim 9 in which each weight has a number of bits w, and the N-way weighted round robin for each virtual output queue is implemented by an N(2^w-1)- way unweighted round robin using a request vector list constructed by interleaving N words of (2^w-1) bits each, each word corresponding to a respective input queue and having a number of "1" s determined by the entry of the ingress port table for that input queue and that virtual output queue.
An apparatus according to claim 10 in which the request vector list is separated into a plurality of round robin blocks, each corresponding to a respective input queue, each ingress multiplexer being arranged to preform a first round robin process independently within each block, and a second round robin process to make a selection among the blocks.
A apparatus according to any of claims 8 to 11 further comprising an external source unit arranged to program the ingress port table (27; 77), the egress port table (28; 78) and the central allocation table (29).
An apparatus according to claim 12 in which the external source unit is arranged to operate using parameters characterizing the length of the virtual output queue and the urgency of the virtual output queue.
An apparatus according to claim 13 in which the external source unit is arranged to operate using a set of sensitivities relating to the length, urgency and pseudo-static bandwidth allocation.